Latest posts of: clyfish

Hello,

I optimized EndianSwap in scrypt130511.cl a little with v_bfi_b32 instruction today.
For me it improved my hashrate from 408 kh/s to 415 kh/s on average.

Search EndianSwap in scrypt130511.cl, and replace

Code:

#define EndianSwap(n) (rotl(n & ES[0], 24U)|rotl(n & ES[1], 8U))

With

Code:

#define EndianSwap(n) (Ch(ES[0], rotl(n, 8U), rotl(n, 24U)))

Then delete *.bin, and restart cgminer.

Before optimization, EndianSwap compiles to:

Code:

v_and_b32     v2, 0x00ff00ff, v1
v_and_b32     v1, 0xff00ff00, v1
v_alignbit_b32  v2, v2, v2, 8
v_alignbit_b32  v1, v1, v1, 24
v_or_b32      v1, v2, v1

After:

Code:

v_alignbit_b32  v2, v1, v1, 24
v_alignbit_b32  v1, v1, v1, 8
s_mov_b32     s0, 0x00ff00ff
v_bfi_b32     v1, s0, v2, v1

From 5 VALUs to 3 VALUs and 1 SALU.

BTW, you can try another optimization I found two weeks ago, It also speeds up my hashrate by about 5kh/s.
Search "i<LOOKUP_GAP" and add "#pragma unroll" before the line.

Code:

#pragma unroll
		for(uint i=0; i<LOOKUP_GAP; ++i) 
			salsa(X);

If you like my work, please donate.

BTC Donate: 1C1Dzhe9V8qwZvCMCTLSu5p4D582rgpBGR
LTC Donate: LVmUpF9eP3YwcZ5r8LLMiK74YfVk7oJpCB
DOGE Donate: D5AQ1y7ukUzo1R6iJxYwbCRD9tgvi3mgws